In the last few years, machine learning has become one of the most important and prolific IT and artificial intelligence branches. It's not surprising that its applications are becoming more widespread day by day in every business sector, always with new and more powerful tools and results. Open source, production-ready frameworks, together with hundreds of papers published every month, are contributing to one of the most pervasive democratization processes in IT history. But why is machine learning so important and valuable?
Since time immemorial, human beings have built tools and machines to simplify their work and reduce the overall effort needed to complete many different tasks. Even without knowing any physical law, they invented levers (formally described for the first time by Archimedes), instruments, and more complex machines to carry out longer and more sophisticated procedures. Hammering a nail became easier and more painless thanks to a simple trick, so did moving heavy stones or wood using a cart. But, what's the difference between these two examples? Even if the latter is still a simple machine, its complexity allows a person to carry out a composite task without thinking about each step. Some fundamental mechanical laws play a primary role in allowing a horizontal force to contrast gravity efficiently, but neither human beings nor horses or oxen knew anything about them. The primitive people simply observed how a genial trick (the wheel) could improve their lives.
The lesson we've learned is that a machine is never efficient or trendy without a concrete possibility to use it with pragmatism. A machine is immediately considered useful and destined to be continuously improved if its users can easily understand what tasks can be completed with less effort or completely automatically. In the latter case, some intelligence seems to appear next to cogs, wheels, or axles. So a further step can be added to our evolution list: automatic machines, built (nowadays we'd say programmed) to accomplish specific goals by transforming energy into work. Wind or watermills are some examples of elementary tools able to carry out complete tasks with minimal (compared to a direct activity) human control.
In the following figure, there's a generic representation of a classical system that receives some input values, processes them, and produces output results:
But again, what's the key to the success of a mill? It's not hasty at all to say that human beings have tried to transfer some intelligence into their tools since the dawn of technology. Both the water in a river and the wind show a behavior that we can simply call flowing. They have a lot of energy to give us free of any charge, but a machine should have some awareness to facilitate this process. A wheel can turn around a fixed axle millions of times, but the wind must find a suitable surface to push on. The answer seems obvious, but you should try to think about people without any knowledge or experience; even if implicitly, they started a brand new approach to technology. If you prefer to reserve the word intelligence to more recent results, it's possible to say that the path started with tools, moved first to simple machines and then to smarter ones.
Without further intermediate (but not less important) steps, we can jump into our epoch and change the scope of our discussion. Programmable computers are widespread, flexible, and more and more powerful instruments; moreover, the diffusion of the internet allowed us to share software applications and related information with minimal effort. The word-processing software that I'm using, my email client, a web browser, and many other common tools running on the same machine are all examples of such flexibility. It's undeniable that the IT revolution dramatically changed our lives and sometimes improved our daily jobs, but without machine learning (and all its applications), there are still many tasks that seem far out of computer domain. Spam filtering, Natural Language Processing, visual tracking with a webcam or a smartphone, and predictive analysis are only a few applications that revolutionized human-machine interaction and increased our expectations. In many cases, they transformed our electronic tools into actual cognitive extensions that are changing the way we interact with many daily situations. They achieved this goal by filling the gap between human perception, language, reasoning, and model and artificial instruments.
Here's a schematic representation of an adaptive system:
Such a system isn't based on static or permanent structures (model parameters and architectures) but rather on a continuous ability to adapt its behavior to external signals (datasets or real-time inputs) and, like a human being, to predict the future using uncertain and fragmentary pieces of information.
What does learning exactly mean? Simply, we can say that learning is the ability to change according to external stimuli and remembering most of all previous experiences. So machine learning is an engineering approach that gives maximum importance to every technique that increases or improves the propensity for changing adaptively. A mechanical watch, for example, is an extraordinary artifact, but its structure obeys stationary laws and becomes useless if something external is changed. This ability is peculiar to animals and, in particular, to human beings; according to Darwin's theory, it's also a key success factor for the survival and evolution of all species. Machines, even if they don't evolve autonomously, seem to obey the same law.
Therefore, the main goal of machine learning is to study, engineer, and improve mathematical models which can be trained (once or continuously) with context-related data (provided by a generic environment), to infer the future and to make decisions without complete knowledge of all influencing elements (external factors). In other words, an agent (which is a software entity that receives information from an environment, picks the best action to reach a specific goal, and observes the results of it) adopts a statistical learning approach, trying to determine the right probability distributions and use them to compute the action (value or decision) that is most likely to be successful (with the least error).
I do prefer using the term inference instead of prediction only to avoid the weird (but not so uncommon) idea that machine learning is a sort of modern magic. Moreover, it's possible to introduce a fundamental statement: an algorithm can extrapolate general laws and learn their structure with relatively high precision only if they affect the actual data. So the term prediction can be freely used, but with the same meaning adopted in physics or system theory. Even in the most complex scenarios, such as image classification with convolutional neural networks, every piece of information (geometry, color, peculiar features, contrast, and so on) is already present in the data and the model has to be flexible enough to extract and learn it permanently.
In the next sections, there's a brief description of some common approaches to machine learning. Mathematical models, algorithms, and practical examples will be discussed in later chapters.
A supervised scenario is characterized by the concept of a teacher or supervisor, whose main task is to provide the agent with a precise measure of its error (directly comparable with output values). With actual algorithms, this function is provided by a training set made up of couples (input and expected output). Starting from this information, the agent can correct its parameters so as to reduce the magnitude of a global loss function. After each iteration, if the algorithm is flexible enough and data elements are coherent, the overall accuracy increases and the difference between the predicted and expected value becomes close to zero. Of course, in a supervised scenario, the goal is training a system that must also work with samples never seen before. So, it's necessary to allow the model to develop a generalization ability and avoid a common problem called overfitting, which causes an overlearning due to an excessive capacity (we're going to discuss this in more detail in the next chapters, however we can say that one of the main effects of such a problem is the ability to predict correctly only the samples used for training, while the error for the remaining ones is always very high).
In the following figure, a few training points are marked with circles and the thin blue line represents a perfect generalization (in this case, the connection is a simple segment):
Two different models are trained with the same datasets (corresponding to the two larger lines). The former is unacceptable because it cannot generalize and capture the fastest dynamics (in terms of frequency), while the latter seems a very good compromise between the original trend and a residual ability to generalize correctly in a predictive analysis.
Formally, the previous example is called regression because it's based on continuous output values. Instead, if there is only a discrete number of possible outcomes (called categories), the process becomes a classification. Sometimes, instead of predicting the actual category, it's better to determine its probability distribution. For example, an algorithm can be trained to recognize a handwritten alphabetical letter, so its output is categorical (in English, there'll be 26 allowed symbols). On the other hand, even for human beings, such a process can lead to more than one probable outcome when the visual representation of a letter isn't clear enough to belong to a single category. That means that the actual output is better described by a discrete probability distribution (for example, with 26 continuous values normalized so that they always sum up to 1).
In the following figure, there's an example of classification of elements with two features. The majority of algorithms try to find the best separating hyperplane (in this case, it's a linear problem) by imposing different conditions. However, the goal is always the same: reducing the number of misclassifications and increasing the noise-robustness. For example, look at the triangular point that is closer to the plane (its coordinates are about [5.1 - 3.0]). If the magnitude of the second feature were affected by noise and so the value were quite smaller than 3.0, a slightly higher hyperplane could wrongly classify it. We're going to discuss some powerful techniques to solve these problems in later chapters.
Common supervised learning applications include:
- Predictive analysis based on regression or categorical classification
- Spam detection
- Pattern detection
- Natural Language Processing
- Sentiment analysis
- Automatic image classification
- Automatic sequence processing (for example, music or speech)
This approach is based on the absence of any supervisor and therefore of absolute error measures; it's useful when it's necessary to learn how a set of elements can be grouped (clustered) according to their similarity (or distance measure). For example, looking at the previous figure, a human being can immediately identify two sets without considering the colors or the shapes. In fact, the circular dots (as well as the triangular ones) determine a coherent set; it is separate from the other one much more than how its points are internally separated. Using a metaphor, an ideal scenario is a sea with a few islands that can be separated from each other considering only their mutual position and internal cohesion.
In the next figure, each ellipse represents a cluster and all the points inside its area can be labeled in the same way. There are also boundary points (such as the triangles overlapping the circle area) that need a specific criterion (normally a trade-off distance measure) to determine the corresponding cluster. Just as for classification with ambiguities (P and malformed R), a good clustering approach should consider the presence of outliers and treat them so as to increase both the internal coherence (visually, this means picking a subdivision that maximizes the local density) and the separation among clusters.
For example, it's possible to give priority to the distance between a single point and a centroid, or the average distance among points belonging to the same cluster and different ones. In this figure, all boundary triangles are close to each other, so the nearest neighbor is another triangle. However, in real-life problems, there are often boundary areas where there's a partial overlap, meaning that some points have a high degree of uncertainty due to their feature values.
Another interpretation can be expressed using probability distributions. If you look at the ellipses, they represent the area of multivariate Gaussians bound between a minimum and maximum variance. Considering the whole domain, a point (for example, a blue star) could potentially belong to all clusters, but the probability given by the first one (lower-left corner) is the highest, and so this determines the membership. Once the variance and mean (in other words, the shape) of all Gaussians become stable, each boundary point is automatically captured by a single Gaussian distribution (except in the case of equal probabilities). Technically, we say that such an approach maximizes the likelihood of a Gaussian mixture given a certain dataset. This is a very important statistical learning concept that spans many different applications, so it will be examined in more depth in the next chapter. Moreover, we're going to discuss some common clustering methodologies, considering both strong and weak points and comparing their performances for various test distributions.
Other important techniques involve the usage of both labeled and unlabeled data. This approach is therefore called semi-supervised and can be adopted when it's necessary to categorize a large amount of data with a few complete (labeled) examples or when there's the need to impose some constraints to a clustering algorithm (for example, assigning some elements to a specific cluster or excluding others).
Commons unsupervised applications include:
- Object segmentation (for example, users, products, movies, songs, and so on)
- Similarity detection
- Automatic labeling
Even if there are no actual supervisors, reinforcement learning is also based on feedback provided by the environment. However, in this case, the information is more qualitative and doesn't help the agent in determining a precise measure of its error. In reinforcement learning, this feedback is usually called reward (sometimes, a negative one is defined as a penalty) and it's useful to understand whether a certain action performed in a state is positive or not. The sequence of most useful actions is a policy that the agent has to learn, so to be able to make always the best decision in terms of the highest immediate and cumulative reward. In other words, an action can also be imperfect, but in terms of a global policy it has to offer the highest total reward. This concept is based on the idea that a rational agent always pursues the objectives that can increase his/her wealth. The ability to see over a distant horizon is a distinction mark for advanced agents, while short-sighted ones are often unable to correctly evaluate the consequences of their immediate actions and so their strategies are always sub-optimal.
Reinforcement learning is particularly efficient when the environment is not completely deterministic, when it's often very dynamic, and when it's impossible to have a precise error measure. During the last few years, many classical algorithms have been applied to deep neural networks to learn the best policy for playing Atari video games and to teach an agent how to associate the right action with an input representing the state (usually a screenshot or a memory dump).
In the following figure, there's a schematic representation of a deep neural network trained to play a famous Atari game. As input, there are one or more subsequent screenshots (this can often be enough to capture the temporal dynamics as well). They are processed using different layers (discussed briefly later) to produce an output that represents the policy for a specific state transition. After applying this policy, the game produces a feedback (as a reward-penalty), and this result is used to refine the output until it becomes stable (so the states are correctly recognized and the suggested action is always the best one) and the total reward overcomes a predefined threshold.
We're going to discuss some examples of reinforcement learning in the chapter dedicated to introducing deep learning and TensorFlow.
During the last few years, thanks to more powerful and cheaper computers, many researchers started adopting complex (deep) neural architectures to achieve goals there were unimaginable only two decades ago. Since 1957, when Rosenblatt invented the first perceptron, interest in neural networks has grown more and more. However, many limitations (concerning memory and CPU speed) prevented massive research and hid lots of potential applications of these kinds of algorithms.
In the last decade, many researchers started training bigger and bigger models, built with several different layers (that's why this approach is called deep learning), to solve new challenging problems. The availability of cheap and fast computers allowed them to get results in acceptable timeframes and to use very large datasets (made up of images, texts, and animations). This effort led to impressive results, in particular for classification based on photo elements and real-time intelligent interaction using reinforcement learning.
The idea behind these techniques is to create algorithms that work like a brain and many important advancements in this field have been achieved thanks to the contribution of neurosciences and cognitive psychology. In particular, there's a growing interest in pattern recognition and associative memories whose structure and functioning are similar to what happens in the neocortex. Such an approach also allows simpler algorithms called model-free; these aren't based on any mathematical-physical formulation of a particular problem but rather on generic learning techniques and repeating experiences.
Of course, testing different architectures and optimization algorithms is quite simpler (and it can be done with parallel processing) than defining a complex model (which is also more difficult to adapt to different contexts). Moreover, deep learning showed better performance than other approaches, even without a context-based model. This suggests that in many cases, it's better to have a less precise decision made with uncertainty than a precise one determined by the output of a very complex model (often not so fast). For animals, this is often a matter of life and death, and if they succeed, it is thanks to an implicit renounce of some precision.
Common deep learning applications include:
- Image classification
- Real-time visual tracking
- Autonomous car driving
- Logistic optimization
- Speech recognition
Many of these problems can also be solved using classic approaches, sometimes much more complex, but deep learning outperformed them all. Moreover, it allowed extending their application to contexts initially considered extremely complex, such as autonomous cars or real-time visual object identification.
This book covers in detail only some classical algorithms; however, there are many resources that can be read both as an introduction and for a more advanced insight.
Many interesting results have been achieved by the Google DeepMind team (https://deepmind.com) and I suggest you visit their website to learn about their latest research and goals.
Another area that can be exploited using machine learning is big data. After the first release of Apache Hadoop, which implemented an efficient MapReduce algorithm, the amount of information managed in different business contexts grew exponentially. At the same time, the opportunity to use it for machine learning purposes arose and several applications such as mass collaborative filtering became reality.
Imagine an online store with a million users and only one thousand products. Consider a matrix where each user is associated with every product by an implicit or explicit ranking. This matrix will contain 1,000,000 x 1,000 cells, and even if the number of products is very limited, any operation performed on it will be slow and memory-consuming. Instead, using a cluster, together with parallel algorithms, such a problem disappears and operations with higher dimensionality can be carried out in a very short time.
Think about training an image classifier with a million samples. A single instance needs to iterate several times, processing small batches of pictures. Even if this problem can be performed using a streaming approach (with a limited amount of memory), it's not surprising to wait even for a few days before the model begins to perform well. Adopting a big data approach instead, it's possible to asynchronously train several local models, periodically share the updates, and re-synchronize them all with a master model. This technique has also been exploited to solve some reinforcement learning problems, where many agents (often managed by different threads) played the same game, providing their periodical contribute to a global intelligence.
Not every machine learning problem is suitable for big data, and not all big datasets are really useful when training models. However, their conjunction in particular situations can drive to extraordinary results by removing many limitations that often affect smaller scenarios.
In the chapter dedicated to recommendation systems, we're going to discuss how to implement collaborative filtering using Apache Spark. The same framework will be also adopted for an example of Naive Bayes classification.
If you want to know more about the whole Hadoop ecosystem, visit http://hadoop.apache.org. Apache Mahout (http://mahout.apache.org) is a dedicated machine learning framework and Spark (http://spark.apache.org), one the fastest computational engines, has a module called MLib that implements many common algorithms that benefit from parallel processing.
An excellent introduction to artificial intelligence can be found in the first few chapters of Russel S., Norvig P., Artificial Intelligence: A Modern Approach, Pearson. In the second volume, there's also a very extensive discussion on statistical learning in many different contexts. A complete book on deep learning is Goodfellow I., Bengio Y., Courville A., Deep Learning, The MIT Press. If you would like to learn more about how the neocortex works, a simple but stunning introduction is present in Kurzweil R., How to Create a Mind, Duckworth Overlook. A comprehensive introduction to the Python programming language can be found in Lutz M., Learning Python, O'Reilly.
In this chapter, we introduced the concept of adaptive systems; they can learn from their experiences and modify their behavior in order to maximize the possibility of reaching a specific goal. Machine learning is the name given to a set of techniques that allow implementing adaptive algorithms to make predictions and to auto-organize input data according to their common features.
The three main learning strategies are supervised, unsupervised, and reinforcement. The first one assumes the presence of a teacher that provides a precise feedback on errors. The algorithm can hence compare its output with the right one and correct its parameters accordingly. In an unsupervised scenario, there are no external teachers, so everything is learned directly from the data. An algorithm will try to find out all features common to a group of elements to be able to associate new samples with the right cluster. Examples of the former type are provided by all the automatic classifications of objects into a specific category according to some known features, while common applications of unsupervised learning are the automatic groupings of items with a subsequent labeling or processing. The third kind of learning is similar to supervised, but it receives only an environmental feedback about the quality of its actions. It doesn't know exactly what is wrong or the magnitude of its error but receives generic information that helps it in deciding whether to continue to adopt a policy or to pick another one.
In the next chapter, we're going to discuss some fundamental elements of machine learning, with particular focus on the mathematical notation and the main definitions that we'll need in all the other chapters. We'll also discuss important statistical learning concepts and some theory about learnability and its limits.